AI Red Teaming Mini-Course: Building Adversarial Examples скачать в хорошем качестве

AI Red Teaming Mini-Course: Building Adversarial Examples

In this week's lesson, I'm taking you back in machine learning history -- learning how to build adversarial examples. Even though today's models are more robust against noise, it's useful to learn how adversarial examples work because: 1. They teach us more about how deep learning works 2. We can still think in adversarial noise paradigms when we build new attacks (and sometimes combine this with actual adversarial attack strategies from many years ago) To follow along, check out the GitHub: https://github.com/kjam/ai-and-ml-sec... Related links from this week's video: Adversarial Turtle: https://www.science.org/content/artic... My CCC talk on adversarial examples in 2017: https://media.ccc.de/v/34c3-8860-deep... FGSM paper from Goodfellow et al. - https://arxiv.org/abs/1412.6572 Adversarial Examples post related to AI/ML Memorization: https://blog.kjamistan.com/adversaria... Early adversarial training paper: https://arxiv.org/abs/1706.06083 Example of using diffusion as denoising: https://arxiv.org/abs/2206.10550